WoFScast: A Machine Learning Model for Predicting Thunderstorms at Watch-to-Warning Scales

Developing AI models that match or exceed the forecast skill of numerical weather prediction (NWP) systems but that can be run much more quickly is a burgeoning area of research. Almost all current AI-NWP models, however, are trained on global ERA5 data, which is too coarse for storm-scale modeling. We have therefore adapted Google's GraphCast framework for training on limited-area, storm-scale domains, then trained on archived data from the Warn-on-Forecast System (WoFS), an 18-member convection-allowing ensemble with forecast output available every 5 minutes. Our initial aim with this "WoFSCast" model is to emulate the WRF-based WoFS forecasts rather than train on the WoFS analyses. We evaluate the WoFSCast predictions using object-based verification, grid-based verification, spatial storm structure assessments, and spectra analysis. The WoFSCast closely emulates the WoFS environment fields and matches 70--80\% of WoFS storms out to 2-h forecast times. Furthermore, when verified against observed storms, WoFSCast produces a similar CSI as WoFS but with a lower frequency bias due to fewer false-alarm storms. Despite using pixel-based MSE loss, we do not observe significant forecast blurring (i.e., loss of detail at smaller scales) of intermittent fields (e.g., composite reflectivity). WoFSCast serves as a proof-of-concept that AI-NWP can be extended to rapidly evolving, small-scale, chaotic phenomena like thunderstorms, even without the use of generative approaches to maintain forecast sharpness.